Documentation
¶
Overview ¶
Package vision turns image bytes into prose a text-only main agent can reason over (master plan §7.4). Used by remote gateways to inline-describe photos and by the vision_describe tool to caption browser screenshots.
We deliberately keep this separate from the providers package: the main agent's Message type is plain string today, and bolting image content blocks onto every provider for a single feature would be a large refactor for marginal gain. Instead a Describer takes bytes in, returns text out, and the caller injects the text wherever it needs it.
Index ¶
Constants ¶
This section is empty.
Variables ¶
This section is empty.
Functions ¶
func MIMEFromBytes ¶
MIMEFromBytes sniffs PNG/JPEG/GIF/WebP magic so callers downloading arbitrary files don't have to guess. Falls back to image/png.
Types ¶
type AnthropicDescriber ¶
type AnthropicDescriber struct {
APIKey string
Model string // e.g. "claude-sonnet-4-5-20250929"
BaseURL string // override for tests; default https://api.anthropic.com
HTTP *http.Client // overridable
MaxTokens int // default 1024
Timeout time.Duration // default 30s
}
AnthropicDescriber is the production Describer. Talks straight to the Messages API rather than going through the providers abstraction because we need image content blocks the providers.Message type doesn't carry yet.
func NewAnthropic ¶
func NewAnthropic(apiKey, model string) *AnthropicDescriber
NewAnthropic returns a describer with sane defaults.
Source Files